Reviewed-by: Kostiantyn Kostiuk <[email protected]> Tested with reproducer from https://gitlab.com/qemu-project/qemu/-/work_items/4584 Behavious is expected
On Wed, Oct 28, 2026 at 12:21 PM Marc-André Lureau < [email protected]> wrote: > Hi > > On Wed, Oct 28, 2026 at 12:08 PM Kostiantyn Kostiuk <[email protected]> > wrote: > > > > > > > > > > > > On Tue, Oct 27, 2026 at 6:06 PM <[email protected]> wrote: > >> > >> From: Marc-André Lureau <[email protected]> > >> > >> Since commit 374f4676788c, the provider's GetSnapshotProperties() > >> returns S_OK for any snapshot ID without filling in the properties. VSS > >> consults every registered provider when a requester looks up a snapshot, > >> so lookups of snapshots owned by other providers now succeed with an > >> all-zero VSS_SNAPSHOT_PROP. Backup software relying on it fails, e.g. > >> Veeam: "Original volume name is not specified for the shadow (shadow > >> volume id: '{00000000-0000-0000-0000-000000000000}')". > >> > >> Simply reverting is not an option: 374f4676788c switched to S_OK because > >> returning VSS_E_OBJECT_NOT_FOUND unconditionally broke qemu-ga's own > >> snapshot sets. The provider must answer for its own snapshots only. It > >> does not create real shadow copies, but it does take part in the > >> snapshot set during qemu-ga fsfreeze cycles. Remember the snapshot IDs > >> received in BeginPrepareSnapshot(), return their required properties, > >> and return VSS_E_OBJECT_NOT_FOUND for anything else. Fill every required > >> VSS_SNAPSHOT_PROP member except the exposed name and path, which are > >> optional. The snapshot device object is an empty string because the > >> host-side snapshot has no guest-visible device. > >> > >> Entries are dropped when VSS deletes the snapshots or the set, or when > >> the set is aborted (AbortSnapshots()). > >> > >> COM+ may hand out several CQGAVssProvider instances (and serve several > >> client processes), so this registry is module-global. It is per-process > >> by design and safe because the COM+ application is hosted by a single NT > >> service (see CreateServiceForApplication() in install.cpp): COM+ forbids > >> pooling and recycling for service-hosted applications, so exactly one > >> surrogate process ever hosts the provider. > >> > >> On the requester side, the snapshot set is created in the persistent > >> VSS_CTX_APP_ROLLBACK context, so VSS keeps it after the freeze/thaw > >> cycle and may keep querying providers about it. The provider registry > >> does not outlive the cycle (requester_thaw() stops the hosting service, > >> and the provider process with it), so later queries would get "not > >> found" for a set VSS still believes exists. Make requester_thaw() delete > >> the snapshot set explicitly at the end of the cycle, whatever its > >> outcome (including BackupComplete() failure and aborted snapshot sets). > >> A failure to delete is only logged, since the guest filesystems have > >> already been thawed. > >> > >> Cc: [email protected] > >> Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/4584 > >> Buglink: https://redhat.atlassian.net/browse/RHEL-250662 > >> Fixes: 374f4676788c ("qga: Fix VSS provider GetSnapshotProperties > return value") > >> Signed-off-by: Marc-André Lureau <[email protected]> > >> --- > >> v2: > >> - fill the required VSS_SNAPSHOT_PROP fields > >> --- > >> qga/vss-win32/install.cpp | 1 + > >> qga/vss-win32/provider.cpp | 200 +++++++++++++++++++++++++++++++++++- > >> qga/vss-win32/requester.cpp | 26 +++++ > >> 3 files changed, 226 insertions(+), 1 deletion(-) > >> > >> diff --git a/qga/vss-win32/install.cpp b/qga/vss-win32/install.cpp > >> index 09d94edd4f53..3b58dc840520 100644 > >> --- a/qga/vss-win32/install.cpp > >> +++ b/qga/vss-win32/install.cpp > >> @@ -331,6 +331,7 @@ STDAPI COMRegister(void) > >> chk(put_Value(pObj, L"Authentication", short(6))); > >> chk(put_Value(pObj, L"AuthenticationCapability", short(2))); > >> chk(put_Value(pObj, L"ImpersonationLevel", short(2))); > >> + /* NT-service-hosted apps have ConcurrentApps=1 enforced */ > >> chk(pApps->SaveChanges(&n)); > >> > >> /* The app should be deleted if something fails after SaveChanges > */ > >> diff --git a/qga/vss-win32/provider.cpp b/qga/vss-win32/provider.cpp > >> index 1c8eb236e57f..bbc30dd57d85 100644 > >> --- a/qga/vss-win32/provider.cpp > >> +++ b/qga/vss-win32/provider.cpp > >> @@ -11,6 +11,11 @@ > >> */ > >> > >> #include "qemu/osdep.h" > >> + > >> +#include <vector> > >> +#include <algorithm> > >> +#include <string> > >> + > >> #include "vss-common.h" > >> #include "vss-debug.h" > >> #ifdef HAVE_VSS_SDK > >> @@ -138,6 +143,170 @@ STDMETHODIMP > CQGAVSSEnumObject::Clone(IVssEnumObject **ppenum) > >> return E_NOTIMPL; > >> } > >> > >> +static VSS_PWSZ qga_vss_strdup(const WCHAR *value) > >> +{ > >> + size_t len = (wcslen(value) + 1) * sizeof(WCHAR); > >> + VSS_PWSZ str = (VSS_PWSZ)CoTaskMemAlloc(len); > >> + > >> + if (str) { > >> + memcpy(str, value, len); > >> + } > >> + return str; > >> +} > >> + > >> +static VSS_PWSZ qga_vss_strdup(const std::wstring &value) > >> +{ > >> + return qga_vss_strdup(value.c_str()); > >> +} > >> + > >> +static VSS_TIMESTAMP qga_vss_timestamp(void) > >> +{ > >> + FILETIME ft; > >> + ULARGE_INTEGER value; > >> + > >> + GetSystemTimeAsFileTime(&ft); > >> + value.LowPart = ft.dwLowDateTime; > >> + value.HighPart = ft.dwHighDateTime; > >> + return value.QuadPart; > >> +} > >> + > >> +/* Registry of the snapshot IDs for this provider */ > >> +class QGAVssSnapshotRegistry > >> +{ > >> +public: > >> + bool Track(VSS_ID set_id, VSS_ID snapshot_id, > >> + VSS_PWSZ volume_name, LONG context); > >> + HRESULT GetProperties(VSS_ID snapshot_id, VSS_SNAPSHOT_PROP *prop) > const; > >> + LONG RemoveSet(VSS_ID set_id); /* returns how many were removed */ > >> + bool Remove(VSS_ID snapshot_id); > >> + > >> +private: > >> + struct Entry { > >> + VSS_ID set_id; > >> + VSS_ID snapshot_id; > >> + std::wstring volume_name; > >> + LONG context; > >> + VSS_TIMESTAMP timestamp; > >> + }; > >> + > >> + std::vector<Entry> m_entries; > >> + mutable SRWLOCK m_lock = SRWLOCK_INIT; > >> +}; > >> + > >> +static QGAVssSnapshotRegistry qga_vss_registry; > >> + > >> +bool QGAVssSnapshotRegistry::Track(VSS_ID set_id, VSS_ID snapshot_id, > >> + VSS_PWSZ volume_name, LONG context) > >> +{ > >> + bool ok; > >> + > >> + AcquireSRWLockExclusive(&m_lock); > >> + try { > >> + m_entries.push_back(Entry { set_id, snapshot_id, volume_name, > >> + context, qga_vss_timestamp() }); > >> + ok = true; > >> + } catch (...) { > >> + ok = false; > >> + } > >> + ReleaseSRWLockExclusive(&m_lock); > >> + return ok; > >> +} > >> + > >> +HRESULT QGAVssSnapshotRegistry::GetProperties(VSS_ID snapshot_id, > >> + VSS_SNAPSHOT_PROP *prop) > const > >> +{ > >> + HRESULT hr = S_OK; > >> + VSS_PWSZ device = NULL, original = NULL; > >> + VSS_PWSZ originating = NULL, service = NULL; > >> + > >> + AcquireSRWLockShared(&m_lock); > >> + auto it = std::find_if(m_entries.begin(), m_entries.end(), > >> + [&](const Entry &e) { > >> + return IsEqualGUID(e.snapshot_id, snapshot_id); > >> + }); > >> + if (it == m_entries.end()) { > >> + hr = VSS_E_OBJECT_NOT_FOUND; > >> + goto out; > >> + } > >> + > >> + { > >> + const Entry &e = *it; > >> + WCHAR machine[256] = L""; > >> + DWORD machine_size = ARRAYSIZE(machine); > >> + > >> + if (!GetComputerNameExW(ComputerNameDnsFullyQualified, machine, > >> + &machine_size)) { > >> + *machine = L'\0'; > >> + } > >> + > >> + device = qga_vss_strdup(L""); > >> + original = qga_vss_strdup(e.volume_name); > >> + originating = qga_vss_strdup(machine); > >> + service = qga_vss_strdup(machine); > >> + if (!device || !original || !originating || !service) { > >> + hr = E_OUTOFMEMORY; > >> + goto out; > >> + } > >> + > >> + memset(prop, 0, sizeof(*prop)); > >> + prop->m_SnapshotId = e.snapshot_id; > >> + prop->m_SnapshotSetId = e.set_id; > >> + prop->m_lSnapshotsCount = std::count_if(m_entries.begin(), > >> + m_entries.end(), [&](const Entry &o) { > >> + return IsEqualGUID(o.set_id, e.set_id); > >> + }); > >> + prop->m_pwszSnapshotDeviceObject = device; > >> + prop->m_pwszOriginalVolumeName = original; > >> + prop->m_pwszOriginatingMachine = originating; > >> + prop->m_pwszServiceMachine = service; > >> + prop->m_ProviderId = g_gProviderId; > >> + prop->m_lSnapshotAttributes = e.context; > >> + prop->m_tsCreationTimestamp = e.timestamp; > >> + prop->m_eStatus = VSS_SS_CREATED; > >> + > >> + device = original = originating = service = NULL; > >> + } > >> + > >> +out: > >> + CoTaskMemFree(device); > >> + CoTaskMemFree(original); > >> + CoTaskMemFree(originating); > >> + CoTaskMemFree(service); > >> + ReleaseSRWLockShared(&m_lock); > >> + return hr; > >> +} > >> + > >> +LONG QGAVssSnapshotRegistry::RemoveSet(VSS_ID set_id) > >> +{ > >> + LONG removed; > >> + > >> + AcquireSRWLockExclusive(&m_lock); > >> + auto it = std::remove_if(m_entries.begin(), m_entries.end(), > >> + [&](const Entry &e) { > >> + return IsEqualGUID(e.set_id, set_id); > >> + }); > >> + removed = m_entries.end() - it; > >> + m_entries.erase(it, m_entries.end()); > >> + ReleaseSRWLockExclusive(&m_lock); > >> + return removed; > >> +} > >> + > >> +bool QGAVssSnapshotRegistry::Remove(VSS_ID snapshot_id) > >> +{ > >> + bool found; > >> + > >> + AcquireSRWLockExclusive(&m_lock); > >> + auto it = std::find_if(m_entries.begin(), m_entries.end(), > >> + [&](const Entry &e) { > >> + return IsEqualGUID(e.snapshot_id, snapshot_id); > >> + }); > >> + found = it != m_entries.end(); > >> + if (found) { > >> + m_entries.erase(it); > >> + } > >> + ReleaseSRWLockExclusive(&m_lock); > >> + return found; > >> +} > >> > >> /* QGAVssProvider */ > >> > >> @@ -266,7 +435,7 @@ STDMETHODIMP CQGAVssProvider::SetContext(LONG > lContext) > >> STDMETHODIMP CQGAVssProvider::GetSnapshotProperties( > >> VSS_ID SnapshotId, VSS_SNAPSHOT_PROP *pProp) > >> { > >> - return S_OK; > >> + return qga_vss_registry.GetProperties(SnapshotId, pProp); > >> } > >> > >> STDMETHODIMP CQGAVssProvider::Query( > >> @@ -286,6 +455,27 @@ STDMETHODIMP CQGAVssProvider::DeleteSnapshots( > >> VSS_ID SourceObjectId, VSS_OBJECT_TYPE eSourceObjectType, > >> BOOL bForceDelete, LONG *plDeletedSnapshots, VSS_ID > *pNondeletedSnapshotID) > >> { > >> + LONG deleted = 0; > >> + > >> + switch (eSourceObjectType) { > >> + case VSS_OBJECT_SNAPSHOT: > >> + if (qga_vss_registry.Remove(SourceObjectId)) { > >> + deleted = 1; > >> + } > >> + break; > >> + case VSS_OBJECT_SNAPSHOT_SET: > >> + deleted = qga_vss_registry.RemoveSet(SourceObjectId); > >> + break; > >> + default: > >> + break; > >> + } > >> + > >> + if (deleted > 0) { > >> + *plDeletedSnapshots = deleted; > >> + *pNondeletedSnapshotID = GUID_NULL; > >> + return S_OK; > >> + } > >> + > >> *plDeletedSnapshots = 0; > >> *pNondeletedSnapshotID = SourceObjectId; > >> return S_OK; > > > > > > return VSS_E_OBJECT_NOT_FOUND in this case looks more correct > > Possibly, but I didn't want to change the existing behaviour here > without good reasons (it was fixed in 2014 by commit > d9e1f574cb6eac0a3a2f97b67d2e7a3ad9c1dc95). > > > > >> > >> @@ -295,6 +485,13 @@ STDMETHODIMP CQGAVssProvider::BeginPrepareSnapshot( > >> VSS_ID SnapshotSetId, VSS_ID SnapshotId, > >> VSS_PWSZ pwszVolumeName, LONG lNewContext) > >> { > >> + if (!pwszVolumeName) { > >> + return E_INVALIDARG; > >> + } > >> + if (!qga_vss_registry.Track(SnapshotSetId, SnapshotId, > >> + pwszVolumeName, lNewContext)) { > >> + return E_OUTOFMEMORY; > >> + } > >> return S_OK; > >> } > >> > >> @@ -411,6 +608,7 @@ STDMETHODIMP > CQGAVssProvider::PostFinalCommitSnapshots(VSS_ID SnapshotSetId) > >> > >> STDMETHODIMP CQGAVssProvider::AbortSnapshots(VSS_ID SnapshotSetId) > >> { > >> + qga_vss_registry.RemoveSet(SnapshotSetId); > >> return S_OK; > >> } > >> > >> diff --git a/qga/vss-win32/requester.cpp b/qga/vss-win32/requester.cpp > >> index 1b0155c691ef..565eefcfb283 100644 > >> --- a/qga/vss-win32/requester.cpp > >> +++ b/qga/vss-win32/requester.cpp > >> @@ -52,6 +52,7 @@ static t_VssFreeSnapshotProperties > pVssFreeSnapshotProperties; > >> static struct QGAVSSContext { > >> IVssBackupComponents *pVssbc; /* VSS requester interface */ > >> IVssAsync *pAsyncSnapshot; /* async info of VSS snapshot > operation */ > >> + VSS_ID idSnapshotSet; /* snapshot set created for this > cycle */ > >> HANDLE hEventFrozen; /* notify fs/writer freeze from > provider */ > >> HANDLE hEventThaw; /* request provider to thaw */ > >> HANDLE hEventTimeout; /* notify timeout in provider */ > >> @@ -116,6 +117,7 @@ static void requester_cleanup(void) > >> vss_ctx.pVssbc->Release(); > >> vss_ctx.pVssbc = NULL; > >> } > >> + vss_ctx.idSnapshotSet = GUID_NULL; > >> vss_ctx.cFrozenVols = 0; > >> qga_debug_end; > >> } > >> @@ -446,6 +448,7 @@ void requester_freeze(int *num_vols, void > *mountpoints, ErrorSet *errset) > >> err_set(errset, hr, "failed to start snapshot set"); > >> goto out; > >> } > >> + vss_ctx.idSnapshotSet = guidSnapshotSet; > >> > >> if (mountpoints) { > >> PWCHAR volume_name_wchar; > >> @@ -680,6 +683,29 @@ void requester_thaw(int *num_vols, void > *mountpints, ErrorSet *errset) > >> err_set(errset, hr, "failed to do snapshot set"); > >> } > >> > >> + /* > >> + * The snapshot set is created in VSS_CTX_APP_ROLLBACK context, so > >> + * it is persistent: if it was committed, VSS keeps it (and keeps > >> + * querying providers about it) until it is explicitly deleted. The > >> + * freeze/thaw cycle is over whatever the outcome above, so delete > >> + * it now. Failure is not fatal here, only log it. > >> + */ > >> + { > >> + LONG lDeletedSnapshots = 0; > >> + VSS_ID idNonDeletedSnapshot; > >> + HRESULT hrDelete; > >> + > >> + hrDelete = > vss_ctx.pVssbc->DeleteSnapshots(vss_ctx.idSnapshotSet, > >> + > VSS_OBJECT_SNAPSHOT_SET, > >> + TRUE, > >> + &lDeletedSnapshots, > >> + > &idNonDeletedSnapshot); > >> + if (FAILED(hrDelete)) { > >> + qga_debug("failed to delete snapshot set after thaw," > >> + " hr=0x%08lx", hrDelete); > >> + } > >> + } > >> + > >> if (err_is_set(errset)) { > >> vss_ctx.pVssbc->AbortBackup(); > >> } > >> -- > >> 2.56.0.116.g6de20f6092dc > >> > >
